Across 321 U.S. occupations and a decade of Bureau of Labor Statistics wage data, AI's measurable effect on employment was zero. Its effect on pay was not. Real wage growth ran about 6.7% slower in high-exposure occupations after 2023 โ and in the lowest pay quartile it ran 10.7% slower. In the highest quartile, nothing: no significant effect at all (Edlich & Slok, Apollo, 2026).
That is your AI productivity gain, located. It did not arrive as headcount reduction. It arrived as raises that never happened, concentrated almost entirely at the bottom of the payroll.
If your 2026 AI business case still models workforce risk as how many roles get eliminated, you are reading the wrong line of the wrong statement.
What Apollo actually measured โ and why "observed adoption" changes the read
Sania Edlich and Torsten Slok of Apollo Global Management built a difference-in-differences model across 321 matched occupations using BLS Occupational Employment and Wage Statistics from 2015 to 2025, with occupation and year fixed effects and standard errors clustered by occupation (Edlich & Slok, Apollo, 2026).
The design choice that matters is how they defined exposure. Nearly every prior estimate โ Felten, Webb, Brynjolfsson-Mitchell-Rock โ scores occupations on theoretical susceptibility derived from O*NET task descriptions. Apollo instead scored exposure on observed usage, drawn from the Anthropic Economic Index. Not which jobs an AI system could plausibly touch, but which jobs it is demonstrably being used for.
That single substitution collapses the population under discussion. High-exposure occupations on the observed measure cover roughly 5.8 million workers โ about 3.7% of the U.S. labor force (Axios, 2026). Three years of theoretical-exposure headlines have been describing a workforce several times larger than the one AI has actually reached.
Inside that smaller population, the two headline coefficients point in opposite directions:
- log(Employment): โ0.0639, standard error 0.0870. Not significant. No detectable job loss.
- log(Real wage): โ0.0667, standard error 0.0181, p < 0.01. Roughly 6.7% slower real wage growth.
The authors re-estimated at exposure cutoffs of 0.4 and 0.6 and the wage result held. Their conservative aggregate lower bound on foregone labor income is $28 billion a year (Edlich & Slok, Apollo, 2026; HR Dive, 2026).
One statistical note worth carrying into any internal discussion: a non-significant employment coefficient is not proof that AI eliminated no jobs. It means this design could not detect an effect, and with that standard error it would have missed a modest one. The wage estimate is the precise one. The employment estimate is the quiet one.
The regressive shape is the actual finding
Split by pre-period wage quartile โ quartiles defined on each occupation's 2015โ2022 mean wage โ the effect is not distributed. It is stacked at the bottom (Edlich & Slok, Apollo, 2026):
- Q1, lowest paid: โ10.7% (p < 0.01)
- Q2: โ5.7% (not significant)
- Q3: โ4.05% (p < 0.1)
- Q4, highest paid: no significant effect
By occupation group, management and professional roles took โ4.1%. Blue-collar work showed no significant effect, which is what you would expect from a technology with limited reach into physical tasks. The service category came in at โ24.3%, but the authors themselves flag it as a small, composition-sensitive subsample. Treat it as a signal about direction, not a number to put on a slide.
The effect was also strongest in smaller occupations โ specialised roles with less task variety to absorb the work AI took over. Narrow jobs have nowhere to redeploy the freed hour internally, so the value shows up in the wage line instead of the workload line.
Now hold that against how most mid-market AI programmes are staffed. The change management, the reskilling budget, the "what does this mean for your career" roadshow โ all of it is aimed at the knowledge and senior layer. That is the quartile where the measured pay effect is zero.
The comp effect is landing where your communication effort is not.
Wage compression is the AI productivity gain nobody approved
There is a reason this went unnoticed for three years. A layoff has an approval gate. It has a business case, a legal review, a severance model, a comms plan, and a name attached to the decision. Wage compression has none of that. It is the emergent sum of a hundred defensible small calls: a role whose scope quietly narrowed, a merit pool allocated on a slightly softer performance rating, a backfill posted at the bottom of the band instead of the middle, a market adjustment deferred one more cycle because the team is "handling volume fine."
No one in that chain decided to capture the AI productivity gain as wage compression. The organisation did it anyway, because compression is what happens by default when a job gets easier and nothing in the compensation process notices.
This also explains why the effect is invisible in your own reporting. You have dashboards for headcount, attrition, and cost per hire. You almost certainly do not have one for real wage growth in AI-exposed roles versus the rest of the payroll. The measurement gap is not an oversight โ it is the reason the mechanism runs unopposed.
Your payroll history stopped being a usable benchmark
The second-order problem is worse than the first. Comp bands at 50โ500 FTE companies are typically anchored on internal history plus a survey cut that lags the market by a year or more. For roles where AI adoption is real, that history now encodes three years of suppressed growth. Benchmarking a 2027 band against it does not correct the compression โ it ratifies it and compounds it for another cycle.
The roles you automated first are precisely the roles for which your own payroll data has stopped being a reliable read of the external market.
What your employees already believe about where the gain went
The distributional fact has a distributional perception attached, and the perception is running ahead of the evidence.
Ipsos polling for the Groundwork Collaborative, fielded in June 2026 with 1,533 U.S. workers, found 51% believe AI's benefits go only or mostly to owners and executives, against 6% who say most or all benefits reach workers (Groundwork Collaborative / Ipsos, 2026). Two-thirds expect AI to make work worse.
Your managers are not equipped to contest that reading. Korn Ferry's Global Total Rewards Pulse Survey, fielded June 2026 across 5,512 organisations in 135 countries, found only 16% of organisations are confident their managers can explain AI-related pay and work changes to employees (Korn Ferry, 2026).
Put the three findings in sequence and the operational risk is obvious. The pay effect is real and regressive. Half the workforce already assumes it. And five out of six companies cannot hold a credible conversation about it. That is not a communications problem downstream of the rollout โ it is the transmission path from a comp decision nobody made to an adoption programme that stalls without anyone naming why.
The counterargument worth taking seriously
Apollo is an asset manager, not a research university, and this is a whitepaper, not a peer-reviewed paper. That is a fair objection and it should be stated plainly rather than managed around.
Three specific limits:
- Venue and incentive. Apollo publishes macro research with a house view. The defence is method, not venue: the specification is conventional, the coefficients and standard errors are published, and the robustness checks are shown. Judge it on the tables.
- Exposure is vendor-derived. Usage-based scoring is a genuine improvement on theoretical exposure, but it comes from one vendor's interaction logs. It is better data, not neutral data.
- Difference-in-differences identifies association under a parallel-trends assumption. If high-exposure occupations were already on a slower wage path for unrelated reasons before 2023, part of the estimate is that pre-existing divergence rather than AI.
None of this is settled, and no mid-market breakout exists โ the application to a 200-person company is inference, not measurement.
But consider what the objection has to buy. To dismiss the finding you must argue that the wage effect is spurious and that it is spurious in a way that happens to concentrate ten times harder in the bottom quartile than the top. The more defensible position is that the magnitude is uncertain and the shape is informative โ and the shape is what changes your decisions.
What to do this quarter
Merge the AI business case and the comp-band review into one document
They are now the same exercise. Any AI initiative that changes what a role does has a compensation consequence, and reviewing them in separate meetings on separate calendars is how the compression happened in the first place. One document, one owner, one review date.
Run the exposure-versus-merit cut on your own payroll
This takes an afternoon, not a project. Tag the roles where AI tooling has materially changed the work. Compare their real wage growth since 2023 โ nominal increases minus inflation โ against the rest of the payroll. If the AI-exposed roles have run behind, you have reproduced the Apollo result inside your own company, and you now have the only version of this evidence that will actually move your leadership team.
Re-benchmark externally, bottom quartile first
For automated roles, go to current external market data rather than internal history. Start at the bottom of the payroll, which is both where the measured effect is largest and where mid-market turnover is already most expensive per unit of tenure.
Script the pay-and-scope conversation before the next rollout
Given the 16% figure, assume your managers cannot currently do this. Write the ten-minute conversation โ what the tool changes about scope, what it does not change about the band, what the progression path is โ and rehearse it with the managers who will deliver it. It is the cheapest intervention in this entire article and almost nobody has funded it.
The decision on your desk
The last three years of AI workforce planning were built around a question โ how many jobs go away โ that the best available observed-adoption evidence answers with "not detectably any." Meanwhile the effect that was measurable, significant, and regressive went straight into the comp cycle, unexamined, because no dashboard in the mid-market is pointed at it. The AI productivity gain was never lost. It was reallocated, quietly, and not to the people who absorbed the change.
Before your next AI investment review, bring one exhibit: real wage growth since 2023 for your AI-exposed roles, against everyone else, split by pay quartile. If the gap is there, you have not been running an efficiency programme. You have been running a pay cut, and you did not approve it.
